08.10.2016 Views

Foundations of Data Science

2dLYwbK

2dLYwbK

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Setting λ = ln(1 + δ)<br />

Prob ( s > (1 + δ)m ) ≤ ( e − ln(1+δ)) (1+δ)m<br />

∏<br />

n e p(eln(1+δ) −1)<br />

i=1<br />

( ) (1+δ)m 1 ∏ n<br />

≤<br />

1 + δ<br />

i=1<br />

( ) (1+δ)m 1<br />

≤<br />

e npδ<br />

(1 + δ)<br />

( ) m<br />

e δ<br />

≤<br />

.<br />

(1 + δ) (1+δ)<br />

e pδ<br />

To simplify the bound <strong>of</strong> Theorem 12.3, observe that<br />

Therefore<br />

(1 + δ) ln (1 + δ) = δ + δ2<br />

2 − δ3<br />

6 + δ4<br />

12 − · · · .<br />

(1 + δ) (1+δ) = e δ+ δ2 2 − δ3 6 + δ4<br />

12 −···<br />

and hence<br />

e δ<br />

(1+δ) (1+δ)<br />

= e − δ2 2 + δ3 6 −···.<br />

Thus, the bound simplifies to<br />

Prob ( s < (1 + δ) m ) ≤ e − δ2 2 m+ δ3 6 m−···.<br />

For small δ the probability drops exponentially with δ 2 .<br />

When δ is large another simplification is possible. First<br />

Prob ( s > (1 + δ) m ) ( ) m ( )<br />

e δ<br />

(1+δ)m e<br />

≤<br />

≤<br />

(1 + δ) (1+δ) 1 + δ<br />

If δ > 2e − 1, substituting 2e − 1 for δ in the denominator yields<br />

Prob(s > (1 + δ) m) ≤ 2 −(1+δ)m .<br />

Theorem 12.3 gives a bound on the probability <strong>of</strong> the sum being greater than the<br />

mean. We now bound the probability that the sum will be less than its mean.<br />

399

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!